Probabilistic clock synchronization in large distributed systems

Clock synchronization within a distributed system is a problem that has been studied extensively in recent years. Of the many solutions proposed thus far, probabilistic synchronization algorithms provide arguably the best compromise between tightness of synchronization and overhead imposed on the system. The main drawbacks of probabilistic algorithms are the requirement of a master/slave organization of clocks, and the relatively high number of synchronization messages that must be sent. These two drawbacks can make them unsuitable for use in large distributed systems. In this paper, we propose a synchronization algorithm that does not use master/slave clocks and reduces the number of synchronization messages needed. The nodes of the system are divided into a number of overlapping groups. Within a group, each node uses one of two probabilistic techniques to estimate the values of other clocks in the group, and uses an interactive convergence algorithm on the resulting estimates to adjust its local clock. Groups are selected so that the maximum skew between any two group members is bounded